AWS re:Invent 2025: Nova 2, Autonomous Agents and the Era of Agentic AI

AWS re:Invent 2025, Amazon’s annual global cloud computing conference, took place from 1st to 5th December in Las Vegas. I had the honour of attending for the first time thanks to the AWS Travel Grant, a program that funds the participation of AWS User Group Leaders from around the world.

As a Solutions Architect at Devoteam, AWS Community Builder, and co-organiser of the AWS User Group Paris, this week represented a unique opportunity to discover the innovations that will shape the solutions we’ll implement for our clients tomorrow.

Going to re:Invent, I had several objectives:

• Discover AWS service innovations
• Understand agentic AI trends. Learn from customer case studies with innovative solutions to better advise and support our clients
• Meet and physically exchange with the AWS community

In this article, you’ll read:

• Day 1 — Monday: GameDay & Community Mixer
• Day 2 — Tuesday: Matt Garman Keynote
• Day 3 — Wednesday: Swami Sivasubramanian Keynote
• Day 4 — Thursday: Infrastructure & Closing Keynote
• Conclusion

Day 1 — Monday: GameDay & Community Mixer

GameDay Cloud Architecture ft. Datadog

The week started strong with the Cloud Architecture GameDay, an intense technical competition in partnership with Datadog. As a team with François, we competed against nearly 60 teams on complex cloud scenarios. This gameday required quick analysis and advanced debugging thanks to the integration of the Datadog platform in our environment.

Result: first place! 🥇

Expo & Community Mixer

This expo at the Venetian brings together hundreds of AWS partners presenting their solutions. It’s always interesting to chat with these partners. It allows you to discover the needs and difficulties identified at AWS customers, and how they address them through their solution.

The Community Mixer in the evening allowed us to meet Community Builders, User Group Leaders, and AWS Heroes from all over the world. The exchanges are truly valuable for understanding how other communities organise and innovate in a caring and fun setting!

Day 2 — Tuesday: Matt Garman Keynote

The keynote by Matt Garman, CEO of AWS, set the tone for this edition. 2025 marks the advent of autonomous AI agents. We’re moving from assistants that respond to prompts to agents that plan, act, and adapt.

Here are the main announcements from the keynote:

Amazon Nova 2: The New Generation of Amazon LLM Models

AWS unveiled its Amazon Nova 2 model family. Designed to promise the best price-performance ratio on the market, it underscores AWS’s determination to catch up with its leading competitors (Anthropic, OpenAI, Google):

• Nova 2 Lite (GA): Economical and fast model for reasoning and code generation. Ideal for chatbots and document processing
• Nova 2 Pro (in Preview): Designed for complex workloads requiring significant reasoning capacity
• Nova 2 Sonic (GA): Real-time speech-to-speech model with support for multiple languages and dynamic voice control
• Nova 2 Omni (in preview): Announced as the first unified multimodal model, accepting text, audio, video, and images as input and providing images and text as output

Another important announcement, Nova Forge!

This service allows customers to customise Amazon’s Nova models to develop deep expertise or adapt them to their specific business context by integrating their data, processes, and particular needs.

The principle consists of providing access to different Nova checkpoints (saved states of the model at specific moments in its training), at various stages, and not just the final, frozen model, as we do when we “fine-tune” an open-weight model.

This then allows us to do “data mixing”. Data mixing consists of injecting general data to keep the general capabilities of the LLM while specialising it in the customer’s domain.

As a result, we obtain what’s called a “frontier” model. It integrates the depth of a business domain, or even a company’s use case.

However, the service is not intended to be accessible to all customers. Pricing is mentioned around $100,000 per year!

Frontier Agents: AI Agents That Work Autonomously

Also, a major announcement: the introduction of a new class of AI agents capable of working in the background with complete autonomy.

Kiro Autonomous Agent (Preview)

Directly integrated into the IDE developed by AWS teams, we now have access to a virtual developer who can work autonomously on development tasks. This is currently only available for Kiro Pro, Pro+, and Power users.

It has very interesting features:

• The agent works asynchronously and autonomously on tasks that can be assigned to it on a GitHub repo. For example, it can currently handle up to 10 tasks in parallel.
• It has a persistent context. This means it learns from the team’s different PRs and the feedback given to it to remember them when processing subsequent tasks.
• Another notable feature is multi-repo management. Imagine you want to update dependencies across fifteen repos. This agent will be able to handle all these updates across different repos in a single request!
• The agent never automatically merges PRs. At the end of each task, it opens a PR for review. You and your team decide to merge it (Human in the loop).
• Integration with tools like Jira, Confluence, GitLab, GitHub, Teams, and Slack.
• The agent can also access the Internet to do web searches and retrieve the latest information.

AWS Security Agent (free during Preview)

This autonomous agent acts as a virtual security engineer. It proactively secures applications throughout their development lifecycle. Among its features:

• Analyses application specifications, architecture documents, and technical designs to identify security risks from the planning phase. The agent then compares all these documents to AWS best practices and a company’s security requirements.
• Automatically analyses PRs against company requirements and the most known vulnerabilities (SQL injection, missing input validation, etc.). It provides remediation advice directly in the GitHub workflow for developers.
• No more annual or quarterly pentests that can stress teams. The agent executes penetration tests on demand with multi-step attack scenarios that can be customised. It creates an attack plan adapted to the application and adapts based on its findings.
• The agent can analyse any application, whether in AWS (via private and public endpoints), on-premise, hybrid, or even in other clouds.

AWS DevOps Agents (Preview)

This autonomous agent acts as a virtual teammate within your DevOps team. It helps resolve and proactively prevent incidents, while continuously improving the reliability and performance of your applications. It has very interesting features.

The agent does two fundamental things. First, it allows faster incident response by analysing metrics, logs, and traces related to a deployment to generate a root cause analysis (RCA) and provide mitigation steps to resolve the incident (MTTR reduction). Second, it helps prevent incidents before they occur by regularly scanning previous incidents, identifying underlying issues, and making directly actionable fix recommendations in 4 areas: observability, infrastructure optimisation, deployment pipeline improvement, and application resilience.

It works like a real team member. Indeed, it can respond to tickets and alarms from ServiceNow, PagerDuty, etc. The agent shares what it finds by creating tickets or sending messages on a Slack or Teams channel. It can also collaborate with other agents like Dynatrace Davis.

Telemetry, Github integration, graph topology

It’s also a telemetry expert. It can connect to observability platforms (Datadog, Splunk, Dynatrace) and others like Grafana and Prometheus via MCP servers. The agent queries metrics, logs, and traces to create RCAs. It recommends improvements in the observability posture implemented in the company.

The agent integrates with GitHub and GitLab to identify incidents related to code deployment and help with rollbacks. It also gives recommendations to improve your deployment pipelines.

The agent builds and maintains a detailed graph topology of your application to inform and guide its tasks. It can also learn the company’s practices and processes from your documentation, runbooks, etc. You also have the option to ask an AWS support engineer to analyse what the agent found.

Amazon Bedrock AgentCore

I’m working more and more with this service, so naturally, this part of the keynote caught my interest.

Amazon Bedrock AgentCore is the platform for building and deploying AI agents at scale. With over 2 million SDK downloads in just 5 months, it supports any agentic framework (CrewAI, LangGraph, LlamaIndex, OpenAI Agents SDK, Strands Agents) and any LLM.

New Features Announced

• Policy (Preview): Define limits on agent actions in natural language that are then translated into Cedar. For example: “an agent that can make refunds up to €100. Beyond this threshold, it must notify a human”.

• Evaluations (Preview): Evaluators to monitor agent behaviour through a score based on metrics on which alarms and alerts can be defined.

• Episodic Memory: Agents learn from past interactions and memorise the feedback given to them to be more accurate and relevant on their next tasks.

Infrastructure

AWS also unveiled important infrastructure advances:

• P6e-GB300 UltraServers: Instances with NVIDIA GB300 NVL72 GPUs, the most advanced GPU architecture in Amazon EC2, very interesting for companies that need to do inference on models with trillions of parameters
• Trainium 3 UltraServers (GA): With a server brought on stage, it brings up to 144 Trainium 3 chips in 3nm, offering 4.4x more compute performance and 4x more energy efficiency than the previous generation.

• Trainium 4 (Preview) — Already in development, promising even more performance and bandwidth
• AWS AI Factories: AWS offers dedicated AI infrastructure deployable in customers’ datacenters, combining NVIDIA GPUs, Trainium chips, and AWS services like Bedrock and SageMaker. Particularly to meet the security and sovereignty requirements faced by some companies.

Modernisation & Other Announcements

AWS Transform integrates new agentic capabilities to modernise any legacy code, including Windows stacks (.NET, SQL Server..), VMware systems, and mainframes.

At the end of the keynote, Matt Garman set himself the challenge of announcing 25 more in 10 minutes with the appearance of a timer on stage!

Among the announcements, I mainly noted:

• Amazon S3: Maximum object size goes from 5 TB to 50 TB, and Batch Operations become 10x faster.
• Lambda Durable Functions: You can now build multi-step applications and AI workflows directly with the Lambda service, without having to use Step Functions. Durable functions automatically checkpoint during their progress, can pause execution for up to a year during long tasks, and can resume after failures. This can be very useful for workflows that require human approvals (human-in-the-loop), LLM call chains, or business processes like customer onboarding or order processing.
• Database Savings Plans: Important announcement for FinOps teams. A new Savings Plan that offers up to 35% savings with a one-year usage commitment
• GuardDuty Extended Threat Detection: This flagship service enables machine learning-based threat detection. It now covers containerised applications deployed with ECS at no additional cost. Already integrated on EKS, AWS completes integration with its container orchestration services.

Community Mixer EMEA and Agentic AI Sessions

At the EMEA Community Mixer, I was able to meet, see again, and discuss with members of the European community. We were able to debrief on the different announcements from the morning.

Then I had the opportunity to attend two sessions:

• Fine-tuning LLMs for Multi-Agent Orchestration (Cosine AI Case Study). I learned best practices for implementing complex multi-agent systems through feedback from the CEO of Cosine AI. The session also explored implementing different techniques to refine the accuracy of models integrated into agents, and also the importance of the model used by the orchestration agent.

• Architecting the future of applications. Several companies presented concrete use cases explaining how agentic AI is transforming their business. With also discussions around governance, reliability, and cost optimisation at scale.

Day 3 — Wednesday: Swami Sivasubramanian Keynote

The day started with the keynote by Swami Sivasubramanian, VP Agentic AI at AWS. The emphasis was on simplifying the creation, deployment, and production of autonomous AI agents using tools such as Strands SDK, Bedrock AgentCore, and the new Frontier Agents.

Amazon Bedrock: Simplified Customisation

• Reinforcement Fine-Tuning. This improves model accuracy by an average of 66% using feedback rather than huge labelled datasets. Customers can create smaller, faster, and cheaper models that perform better on their specific use cases. Basically, get the quality of a large model at the price of a small one.

Reinforcement Fine-Tuning

• 18 new open-weight models. Including Mistral Large 3, Qwen, MiniMax, OpenAI gpt-oss, DeepSeek-V3.1, bringing the total to around 100 serverless models

Strands SDK & SageMaker

Strands Agents SDK

AWS’s open-source framework for building AI agents has surpassed 5 million downloads. Major new announcement: TypeScript support is now available in preview. We can now choose between Python and TypeScript to develop our agents.

Amazon Nova Act

Nova Act becomes a standalone service, designed to “build, deploy, and manage fleets of reliable agents to automate production UI workflows”, AWS announces the goal of exceeding 90% success rate on repetitive tasks at scale. The service uses a custom Nova 2 Lite model as its base, which has been trained to act on web browsers by navigating websites. It can click buttons, fill out forms… without human intervention! The goal is to allow customers to transform workflows that still require manual actions by automating them.

User Group Leader Meeting

Highlight of the day: meeting with User Group leaders from all over the world. With presentations on the challenges most often encountered when organising events. We’re coming back with excellent ideas to improve AWS User Group meetups in Paris with François.

At the end, we had the surprise of receiving a visit from Jeff Barr, AWS Chief Evangelist. He delivered a very inspiring message about the importance of contributing to communities. His words really touched me, because when I started using AWS and the cloud, the meetings, blog posts, and conferences by other people helped me a great deal in learning and progressing, and this is still the case today. I’m now happy to be able to contribute to the community in turn.

Day 4 — Thursday: Infrastructure & Closing Keynote

Infrastructure Keynote

In this keynote, Dave Brown (Vice President, AWS Compute & ML Services) & Peter DeSantis (SVP, AWS Utility Computing Products) unveiled major announcements. Here are the ones that struck me most.

Graviton5: AWS’s Most Powerful CPU

The new Graviton5 processors: 192 cores per chip, inter-core latency reduced by 33%, L3 cache multiplied by 5. EC2 M9g instances equipped with this chip developed by AWS offer up to 25% more performance compared to Graviton4. Apple, Adobe, Airbnb, Epic Games, and SAP are already using these instances in production.

Amazon S3 Vectors (GA)

This service is now GA with very interesting improvements: up to 2 billion vectors per index, 20 trillion vectors per bucket, 100 ms query latencies, and cost reduction up to 90% compared to vector databases.

Amazon EKS Capabilities

EKS Capabilities now allows using popular tools ArgoCD, ACK, KRO in a fully managed version by AWS. This means the maintenance of these tools runs and is managed in the AWS accounts where the control planes are. This allows customers using EKS Auto Mode to further reduce the operational management of their EKS clusters.

Workshop & Meetup

During the morning, I was able to participate in a hands-on workshop on Strands, MCP, and A2A to better understand best practices in deploying AI agents, architecture patterns, and agent observability.

I was also able to participate in a Meetup, which is a session format that allows exchanging and sharing agent deployment challenges with AWS Solutions Architects and other participants. Very enriching and underrated in my opinion!

In a hallway, I was able to meet Corey Quinn, author of the popular blog Last Week in AWS, which decrypts AWS pricing announcements in a rather sarcastic tone.

Closing Keynote: Werner Vogels and the Renaissance Developer

The closing keynote by Werner Vogels, Amazon’s CTO, will remain memorable. After 14 years of re:Invent keynotes, Werner announced that this AWS re:Invent 2025 edition would be the last. Not that he’s leaving Amazon, but to make room for new voices.

Will AI take my job? Maybe. Will AI make me obsolete? Absolutely not… if you evolve.

Werner then introduced the concept of the “Renaissance Developer”. Like Renaissance thinkers who mastered science, art, and engineering, tomorrow’s developers will need to cultivate five essential qualities:

1. Stay curious and learn from failures by experimenting.
2. Think system, not in isolated components.
3. Communicate: share with peers, don’t neglect the social aspect, continue learning with others through meetups, User Groups, conferences, blogs…
4. Own your work: “The work is yours, not the tool’s”. AI generates code, but we are responsible for decisions about quality, security, and architecture.
5. Become a polymath: It will become even more important to extend your knowledge into other fields.

Werner also emphasised the importance of code review in the AI era:

The review becomes the control point to restore balance. It is where we bring human judgment back into the loop.

Currently working on implementing AI agents for several clients, this message comes at the right time. The tools are evolving rapidly, but our role as architects, our judgment is irreplaceable.

If you want to learn more about his predictions, I highly recommend reading this article.

Or you can simply rewatch the keynote on YouTube, here.

Conclusion

This edition of AWS re:Invent 2025 was, for me, much more than a week filled with announcements.

It highlighted a revolutionary transformation we are witnessing: AI is evolving from the status of assistant to that of autonomous agent, integrated at the heart of cloud architectures. The services presented, from Kiro to Nova Forge through Strands and Bedrock AgentCore, show that AWS is preparing a future where systems will be able to collaborate with teams, not just assist them.

But beyond the announcements, this week also reminded me of the importance of community. The meetings with Community Builders, User Group Leaders, and AWS Heroes, the discussions at EMEA mixers, the feedback from User Group Leaders, not to mention Jeff Barr’s speech, reminded me how much cloud innovation remains first and foremost a human adventure. It’s in these exchanges that we understand what really works in the field.

And that’s precisely where the discourse became nuanced.

A Powerful Ecosystem That’s Growing, But a Pace That Exceeds the Maturity of Some Customers

Analysts are unanimous: AWS’s approach is solid, consistent, and pushes innovation in the right direction. But it’s moving fast… Perhaps too fast for a large part of the market.

Many companies are still at the GenAI prototype stage, while announcements already talk about autonomous agents in production, advanced reasoning, and real-time agentic governance. The risk? Some customers find themselves on the sidelines, lacking the data foundations and governance sufficient to keep pace.

AWS continues to excel in raw capabilities: computing power, chips, scalability, performance, and security. The Nova models themselves are performant but not yet differentiating compared to the competition. On this terrain, AWS bets more on cost efficiency and scalability than on raw performance.

AWS’s Strategy: Master the Entire Chain

Trainium3, Graviton5, Nova Forge, AgentCore, Strands… Everything converges toward a vision: becoming the most economical and most scalable platform for agentic AI in production. An ambitious, consistent vision, but also demanding for companies that want to keep up.

What This Means for Our Clients

This year, more than ever, I come back with a conviction: this transition to autonomous AI can only happen if the foundations are solid.

Companies will need:

• Clean, governed, catalogued data
• Architectures ready for agentic AI
• Integrated pipelines
• And good support to bridge vision and execution.

My Takeaway from AWS re:Invent 2025

This week at re:Invent brought me a lot in a relatively short time. Technically, of course, but also humanly. The exchanges with the community, the workshops, the feedback, the sometimes very frank discussions… all of this reminded me why I love this profession: for continuous learning and sharing!

I’m coming back with lots of ideas to test, things to share, and the desire to transform these innovations into real projects for our clients.